Machine Intelligence - Part 3 of Piero Scaruffi's class "Thinking about Thought" at UC Berkeley (2014), excerpted from http://www.scaruffi.com/nature I keep updating these slides at www.scaruffi.com/ucb.html
1. 1
Thinking about Thought
• Introduction
• Philosophy of Mind
• Cognitive Models
• Machine Intelligence
• Life and Organization
• Ecology
• The Brain
• Dreams and Emotions
• Language
• Modern Physics
• Consciousness
Make it idiot proof and someone will make a better idiot
(One-liner signature file on the internet)
2. 2
Session Three: Machine Intelligence
for Piero Scaruffi's class
"Thinking about Thought"
at UC Berkeley (2014)
Roughly These Chapters of My Book “Nature
of Consciousness”:
3. Machine Intelligence
5. Common Sense: Engineering The Mind
6. Connectionism And Neural Machines
3. 3
Mind and Machines
Is our mind a machine?
Can we build one?
Mathematical models of the mind
Descartes: water fountains
Freud: a hydraulic system
Pavlov: a telephone switchboard
Wiener: a steam engine
Simon: a computer
The computer is the first machine that can be
programmed to perform different tasks
4. 4
A Brief History of Logic
Pythagoras’s' theorem (6th c BC): a relationship between
physical quantities that is both abstract and eternal
Euclides' "Elements" (350 BC), the first system of Logic,
based on just 5 axioms
Aristoteles' "Organon" (4th c BC): syllogisms
William Ockham's "Summa Totius Logicae" (1300 AD) on
how people reason and learn
Francis Bacon’s "Novum Organum" (1620)
Rene' Descartes’ "Discours de la Methode" (1637): the
analytic method over the dialectic method
Gottfried Leibniz ‘s "De Arte Combinatoria" (1676)
Leonhard Euler (1761) how to do symbolic logic with
diagrams
Augustus De Morgan's "The Foundations of Algebra" (1844)
5. 5
A Brief History of Logic
George Boole's "The Laws Of Thought" (1854): the
laws of logic “are” the laws of thought
Propositional logic and predicate logic: true/false!
6. 6
A Brief History of Logic
Axiomatization of Thought:
Gottlob Frege's "Foundations of Arithmetic"
(1884)
Giuseppe Peano's "Arithmetices Principia
Nova Methodo Exposita" (1889)
Bertrand Russell's "Principia Mathematica"
(1903)
7. 7
The Axiomatic Method
• David Hilbert (1900)
– Hilbert’s second problem: prove
the consistency of the axioms of
arithmetic
8. 8
The Axiomatic Method
• David Hilbert (1928)
– Entscheidungsproblem problem: the
mechanical procedure for proving
mathematical theorems
– An algorithm, not a formula
– Mathematics = blind manipulation of symbols
– Formal system = a set of axioms and a set of
inference rules
– Propositions and predicates
– Deduction = exact reasoning
– Logic emancipated from reality by dealing
purely with abstractions
9. 9
The Axiomatic Method
• Paradoxes
– ”I am lying"
– The class of classes that do not belong
to themselves (the barber who shaves
all barbers who do not shave
themselves)
– The omnipotent god
10. 10
The Axiomatic Method
• Kurt Goedel (1931)
– The answer to Hilbert’s second problem
– Any formal system contains an
“undecidable” proposition
– A concept of truth cannot be defined
within a formal system
– Impossible to reduce logic to a
mechanical procedure to prove
theorems (“decision problem”)
11. 11
The Axiomatic Method
• Alfred Tarski (1935)
– Definition of “truth”: a statement is true if it
corresponds to reality (“correspondence
theory of truth”)
– Truth is defined in terms of physical concepts
– Logic grounded back into reality
12. 12
The Axiomatic Method
• Alfred Tarski (1935)
– Base meaning on truth, semantics on logic
(truth-conditional semantics)
– “Interpretation” and “model” of a theory
(“model-theoretic” semantics)
– Theory = a set of formulas.
– Interpretation of a theory = a function that
assigns a meaning (a reference in the real
world) to each formula
– Model for a theory = interpretation that satisfies
all formulas of the theory
– The universe of physical objects is a model for
physics
13. 13
The Axiomatic Method
• Alfred Tarski (1935)
– Build models of the world which yield
interpretations of sentences in that world
– Meaning of a proposition=set of situations in
which it is true
– All semantic concepts are defined in terms of
truth
– Meaning grounded in truth
– Truth can only be relative to something
– “Meta-theory”
14. 14
The Axiomatic Method
• Alan Turing (1936)
– Hilbert’s challenge (1928): an algorithm
capable of solving all the mathematical
problems
– Turing Machine (1936): a machine whose
behavior is determined by a sequence of
symbols and whose behavior determines the
sequence of symbols
– A universal Turing machine (UTM) is a Turing
machine that can simulate an arbitrary Turing
machine
15. 15
The Axiomatic Method
• Alan Turing (1936)
– Computation = the formal manipulation of
symbols through the application of formal
rules
– Turing machine = a machine that is capable
of performing any type of computation
– Turing machine = the algorithm that Hilbert
was looking for
– Hilbert’s program reduced to manipulation of
symbols
– Problem solving = symbol processing
16. 16
The Axiomatic Method
• The Turing machine:
– …an infinite tape (an unlimited
memory capacity)
– … marked out into squares, on each
of which a symbol can be printed…
– The machine can read or write one
symbol at a time
– At any moment there is one symbol in
the machine, the scanned symbol
– The machine can alter the scanned
symbol based on that symbol and on a
table of instructions
– The machine can also move the tape
back and forth
17. 17
The Axiomatic Method
• Alan Turing (1936)
– Universal Turing Machine: a Turing
machine that is able to simulate any other
Turing machine
– The universal Turing machine reads the
description of the specific Turing machine
to be simulated
Turing Machine
18. 18
The Axiomatic Method
• Turing machines in nature: the ribosome, which
translates RNA into proteins
– Genetic alphabet: nucleotides ("bases"): A, C, G, U
– The bases are combined in groups of 3 to form "codons“
– RNA is composed of a string of nucleotides ("bases")
according to certain rules
– There are special carrier molecules ("tRNA") that are
attached to specific aminoacids (proteins)
– The start codon encodes the aminoacid Methionine
– A codon is matched with a specific tRNA
– The new aminoacid is attached to the protein
– The tape then advances 3 bases to the next codon, and the
process repeats
– The protein keeps growing
– When the “stop” codon is encountered, the ribosome
dissociates from the mRNA
19. 19
The Axiomatic Method
• Alan Turing (1936)
– Computable functions = “recursive” functions
– Recursion, the Lambda-calculus, and the Turing
machine are equivalent
– Each predicate is defined by a function, each
function is defined by an algorithm
20. 20
Computer Science
1941: Konrad Zuse's Z3 programmable electronic
computer
1943: Warren McCulloch's and Walter Pitts' binary
neuron
21. 21
Computer Science
• World War II:
– Breaking the Enigma code (Bombe)
– Turing worked at Bletchley Park where the
Colossus was built but it was not a universal
Turing machine (not general purpose)
Replica of the Bombe
22. 22
Computer Science
1945: John Von Neumann's computer architecture
Control unit:
•reads an instruction from memory
•interprets/executes the instruction
•signals the other components what to do
•Separation of instructions and data (although
both are sequences of 0s and 1s)
•Sequential processing
23. 23
Computer Science
• The first non-military computer, ENIAC, by John
Mauchly and Presper Eckert (1946)
24. 24
Computer Science
1947: John Von Neumann's self-
reproducing automata
1. A universal constructor A, which can
construct another automaton according
to instruction I.
2. A copier B, which can make a copy of
the instruction tape I.
3. A controller C, which combines A and B
and allows A to construct a new
automaton according to I, B to copy
instructions from I and attach them to
the newly created automaton
4. An instruction tape describing how to
construct automaton A+B+C
25. 25
Cybernetics
1943: Beginning of Cybernetics
1946: The first Macy Conference on Cybernetics
– John von Neumann (computer science),
– Norbert Wiener, Walter Pitts (mathematics),
– Rafael Lorente de No, Ralph Gerard, etc (neurology),
– Warren McCulloch (neuropsychiatry),
– Gregory Bateson, Margaret Mead (anthropology),
– Heinrich Kluever, Lawrence Kubie, etc (psychology)
– Filmer Northrop (philosophy),
– Lawrence Frank, Paul Lazarsfeld (sociology)
26. 26
Cybernetics
Norbert Wiener (1947)
• Bridge between machines and nature,
between "artificial" systems and natural
systems
• Feedback, by sending back the output as
input, helps control the work of the machine
• A control system is realized by a loop of
action and feedback
• A control system is capable of achieving a
"goal", is capable of "purposeful" behavior
• Living organisms are control systems
vs
27. 27
Cybernetics
• Nicholas Bernstein (1920s): self-regulatory
character of the human nervous system
• Walter Cannon (1930s): Feedback is crucial
for "homeostasis", the process by which an
organism tends to compensate variations in
the environment in order to maintain its
internal stability
• Message
• Noise
• Information
• Entropy = a measure of disorder = a measure
of the lack of information
28. 28
Cybernetics
Norbert Wiener (1947)
• Information is the opposite of entropy:
– “The amount of information in a system is a
measure of its degree of organization, so
the entropy of a system is a measure of its
degree of disorganization"
– "The quantity that we here define as
amount of information is the negative of the
quantity usually defined as entropy in
similar situations."
– “The processes which lose information
are... closely analogous to the processes
which gain entropy. "
29. 29
Cybernetics
• Paradigm shift from the world of continuous
laws to the discrete world of algorithms
• The effect of an algorithm is to turn time’s
continuum into a sequence of discrete
quanta, and, correspondingly, to turn an
analog instrument into a digital instrument
• A watch is the digital equivalent of a
sundial: the sundial marks the time in a
continuous way, the watch advances by
units.
(Ikea)
31. 31
Cybernetics
Cybernetics
• An analog instrument can be precise, and
there is no limit to its precision.
• A digital instrument can only be approximate,
its limit being the smallest magnitude it can
measure
Titan Octane 9308KM01
32. 32
Cybernetics
Ross Ashby (1952)
– Both machines and living beings tend to
change in order to compensate variations in the
environment, so that the combined system is
stable
– The "functioning" of both living beings and
machines depends on feedback processes
– The system self-organizes
– In any isolated system, life and intelligence
inevitably develop
33. 33
Cybernetics
William Powers (1973)
• Living organisms and some machines
are made of hierarchies of control
systems
• Behavior is a backward chain of
behaviors: walking up the hierarchy
one finds out why the system is doing
what it is doing (e.g., it is keeping the
temperature at such a level because
the engine is running at such a speed
because… and so forth).
• The hierarchy is a hierarchy of goals
(goals that have to be achieved in
order to achieve other goals in order
to achieve other goals in order to…)
34. 34
Cybernetics
William Powers (1973)
• Instinctive behavior is the result of the
interaction between control systems that
have internal goals
• In a sense, there is no learning: there is just
the blind functioning of a network of control
systems.
• In another sense, that "is" precisely what we
call "learning": a control system at work
• A hierarchy of control systems can create the
illusion of learning and of intelligence
35. 35
Information Theory
Claude Shannon (1949)
• The entropy of a question is related to the
probability assigned to all the possible
answers to that question
• The amount of information given the
probabilities of its component symbols
• The amount of information is equal to
entropy.
36. 36
Information Theory
Claude Shannon (1949)
– Information
• "Information is a measure of one’s
freedom of choice in selecting a
message. The greater this freedom of
choice, the greater the information, the
greater is the uncertainty that the
message actually selected is some
particular one." (Weaver)
37. 37
Information Theory
Claude Shannon (1949)
– Information and entropy
• Shannon`s "information" is the opposite
of Wiener`s "information".
• Shannon defines information as chaos
(entropy), Wiener and Brillouin define
information as order (negentropy)
• Shannon’s entropy of information is
Brillouin’s negentropy
38. 38
Information Theory
Claude Shannon (1949)
– No interest in the meaning of the message:
• "The fundamental problem of
communication is that of reproducing at
one point exactly or approximately a
message selected at another point.
Frequently the messages have meaning;
that is they refer to or are correlated
according to some system with certain
physical or conceptual entities. These
semantic aspects of communication are
irrelevant to the engineering problem. The
significant aspect is that they are selected
from a set of possible meanings."
39. 39
Information Theory
Entropy
• Sadi Carnot:
– Steam engines cannot exceed a specific
maximum efficiency
– There are no reversible processes in nature
• “I propose to name the quantity S the entropy
of the system, after the Greek word [τροπη
trope], the transformation” (Rudolf Clausius,
1865)
• Clausius: the ratio between absorbed energy
and absolute temperature
40. 40
Information Theory
Entropy
• Ludwig Boltzmann: the number of molecular
degrees of freedom
• Boltzmann’s entropy is nothing more than
Shannon’s entropy applied to equiprobable
microstates.
• Entropy can be understood as “missing”
information
• Entropy measures the amount of disorder in a
physical system; i.e. entropy measures the lack
of information about the structure of the system
Entropy = - Information
41. 41
Information Theory
Entropy
• Leon Brillouin's negentropy principle of
information (1953)
– New information can only be obtained at
the expense of the negentropy of some
other system.
42. 42
Information Theory
Vladimir Kotelnikov (1933)
• The “Nyquist–Shannon” sampling theorem: how to
convert continuous signals (analog domain) into
discrete signals (digital domain) and back without
losing fidelity
43. 43
Information Theory
Andrei Kolmogorov (1963)
• Algorithmic information theory
• Complexity = quantity of information
• Complexity of a system = shortest possible description of
it = the shortest algorithm that can simulate it = the size
of the shortest program that computes it
• Randomness: a random sequence is one that cannot be
compressed any further
• Random number: a number that cannot be computed
(that cannot be generated by any program)
44. 44
Prehistory of Artificial Intelligence
Summary
– David Hilbert (1928)
– "I am lying"
– The omnipotent god
– Kurt Goedel (1931)
– Alfred Tarski (1935)
– Alan Turing (1936)
– Norbert Wiener (1947)
– Claude Shannon and Warren Weaver (1949)
– Entropy = a measure of disorder = a measure of the
lack of information
– Complexity
45. 45
History of Artificial Intelligence
1950: Alan Turing's "Computing Machinery and
Intelligence" (the "Turing Test")
46. 46
History of Artificial Intelligence
The Turing Test (1950)
• A machine can be said to be “intelligent” if it
behaves exactly like a human being
• Hide a human in a room and a machine in
another room and type them questions: if you
cannot find out which one is which based on
their answers, then the machine is intelligent
47. 47
History of Artificial Intelligence
The “Turing point”: a computer can be said to be intelligent
if its answers are indistinguishable from the answers of a
human being
??
48. 48
History of Artificial Intelligence
The fundamental critique to the Turing Test
• The computer cannot (qualitatively) do what the
human brain does because the brain
– does parallel processing rather than sequential
processing
– uses pattern matching rather than binary logic
– is a connectionist network rather than a Turing
machine
49. 49
History of Artificial Intelligence
The Turing Test
• John Searle’s Chinese room (1980)
– Whatever a computer is computing, the
computer does not "know" that it is
computing it
– A computer does not know what it is doing,
therefore “that” is not what it is doing
– Objection: The room + the machine “knows”
50. 50
History of Artificial Intelligence
The Turing Test
• Hubert Dreyfus (1972):
– Experience vs knowledge
– Meaning is contextual
– Novice to expert
– Minds do not use a theory about the everyday world
– Know-how vs know that
• Terry Winograd
– Intelligent systems act, don't think.
– People are “thrown” in the real world
51. 51
History of Artificial Intelligence
The Turing Test
• Rodney Brooks (1986)
– Situated reasoning
– Intelligence cannot be separated from the body.
– Intelligence is not only a process of the brain, it is
embodied in the physical world
– Cognition is grounded in the physical interactions
with the world
– There is no need for a central representation of
the world
– Objection: Brooks’ robots can’t do math
52. 52
History of Artificial Intelligence
The Turing Test
• John Randolph Lucas (1961) & Roger Penrose
– Goedel’s limit: Every formal system (>Arithmetic)
contains a statement that cannot be proved
– Some logical operations are not computable,
nonetheless the human mind can treat them (at
least to prove that they are not computable)
– The human mind is superior to a computing
machine
53. 53
History of Artificial Intelligence
The Turing Test
• John Randolph Lucas (1961) & Roger Penrose
– Objection: a computer can observe the failure of
“another” computer’s formal system
– Goedel’s theorem is about the limitation of the
human mind: a machine that escapes Goedel’s
theorem can exist and can be discovered by
humans, but not built by humans
54. 54
The Turing Test
• What is measured: intelligence, cognition, brain,
mind, or consciousness?
• What is measured: one machine, ..., all machines?
• What is intelligence? What is a brain? What is a
mind? What is life?
• Who is the observer? Who is the judge?
• What is the instrument (instrument = observer)?
• What if a human fails the Turing test?
History of Artificial Intelligence
55. 55
The ultimate Turing Test
• Build a machine that reproduces my brain, neuron by
neuron, synapses by synapses
• Will that machine behave exactly like me?
• If yes, is that machine “me”?
History of Artificial Intelligence
56. 56
History of Artificial Intelligence
1954: Demonstration of a machine-translation
system by Leon Dostert's team at
Georgetown University and Cuthbert
Hurd's team at IBM
1956: Dartmouth conference on Artificial
Intelligence
Artificial Intelligence (1956): the discipline of
building machines that are as intelligent as
humans
57. 57
History of Artificial Intelligence
1956: Allen Newell and Herbert Simon
demonstrate the "Logic Theorist“, the first
A.I. program, that uses “heuristics” (rules of
thumb) and proves 38 of the 52 theorems
in Whitehead’s and Russell’s “Principia
Mathematica”
1957: “General Problem Solver” (1957): a
generalization of the Logic Theorist but
now a model of human cognition
58. 58
History of Artificial Intelligence
1957: Noam Chomsky's "Syntactic Structures"
S stands for Sentence, NP for Noun Phrase, VP for Verb Phrase, Det for Determiner,
Aux for Auxiliary (verb), N for Noun, and V for Verb stem
59. 59
History of Artificial Intelligence
1957: Frank Rosenblatt's Perceptron, the first
artificial neural network
60. 60
History of Artificial Intelligence
1959: John McCarthy's "Programs with Common
Sense" (1949) focuses on knowledge
representation
1959: Arthur Samuel's Checkers, the world's first
self-learning program
1960: Hilary Putnam's Computational Functionalism
(see chapter on “Philosophy of Mind”)
1962: Joseph Engelberger deploys the industrial
robot Unimate at General Motors
61. 61
History of Artificial Intelligence
1963 Irving John Good speculates about
"ultraintelligent machines" (the "singularity")
1964: IBM's "Shoebox" for speech recognition
1965: Ed Feigenbaum's Dendral expert system:
domain-specific knowledge
64. 64
History of Artificial Intelligence
1966: Joe Weizenbaum's Eliza
1968: Peter Toma founds Systran to commercialize machine-
translation systems
65. 65
History of Artificial Intelligence
1969: Marvin Minsky & Samuel Papert's
"Perceptrons" kill neural networks
1969: Stanford Research Institute's Shakey the
Robot
1972: Bruce Buchanan's MYCIN
•a knowledge base
•a patient database
•a consultation/explanation program
•a knowledge acquisition program
Knowledge is organised as a series of IF THEN rules
67. 67
History of Artificial Intelligence
1972: Hubert Dreyfus's "What Computers Can't Do"
1974: Marvin Minsky's Frame (see chapter on
“Cognition”)
1975: Roger Schank's Script (see chapter on
“Cognition”)
1975: John Holland's Genetic Algorithms
1976: Doug Lenat's AM
1979: Cordell Green's system for automatic
programming
1979: Drew McDermott's non-monotonic logic
1979: David Marr's theory of vision
68. 68
History of Artificial Intelligence
1980: John Searle’s "Chinese Room"
1980: Intellicorp, the first major start-up for Artificial
Intelligence
1982: Japan's Fifth Generation Computer Systems
project
69. 69
History of Artificial Intelligence
1982: John Hopfield describes a new generation of neural
networks, based on a simulation of annealing
1983: Geoffrey Hinton's and Terry Sejnowski's Boltzmann
machine for unsupervised learning
1986: Paul Smolensky's Restricted Boltzmann machine
1986: David Rumelhart’s “Parallel Distributed Processing”
Rummelhart network
Neurons arranged in layers, each neuron
linked to neurons of the neighboring
layers, but no links within the same layer
Requires training with supervision
Hopfield networks
Multidirectional data flow
Total integration between input and output
data
All neurons are linked between themselves
Trained with or without supervision
70. 70
History of Artificial Intelligence
Genealogy of Intelligent Machines
Hydraulic machines
Steam engines
Cybernetics
Neural networks
Logic
Hilbert
Turing Machine
Computers
Expert Systems
71. 71
History of Artificial Intelligence
• Valentino Breitenberg’s “vehicles” (1984)
– Vehicle 1: a motor and a sensor
– Vehicle 2: two motors and two sensors
– Increase little by little the circuitry, and these
vehicles seem to acquire not only new skills, but
also a personality.
72. 72
History of Artificial Intelligence
1985: Judea Pearl's "Bayesian Networks"
1987: Chris Langton coins the term "Artificial Life"
1987: Rodney Brooks' robots
1990: Carver Mead describes a neuromorphic
processor
1992: Thomas Ray develops "Tierra", a virtual world
73. 73
History of Artificial Intelligence
1997: IBM's "Deep Blue" chess machine beats the world's
chess champion, Garry Kasparov
2011: IBM's Watson debuts on a tv show
74. 74
History of Artificial Intelligence
2000: Cynthia Breazeal's emotional robot, "Kismet"
2003: Hiroshi Ishiguro's Actroid, a young woman
75. 75
History of Artificial Intelligence
2004: Mark Tilden's biomorphic robot Robosapien
2005: Honda's humanoid robot "Asimo"
76. 76
History of Artificial Intelligence
2005: Boston Dynamics' quadruped robot "BigDog"
2010: Lola Canamero's Nao, a robot that can show its
emotions
2011: Osamu Hasegawa's SOINN-based robot that learns
functions it was not programmed to do
2012: Rodney Brooks' hand programmable robot "Baxter"
77. 77
History of Artificial Intelligence
• Rod Brooks/ Rethink Robotics (2012)
– Vision to locate and grasp objects
– Can be taught to perform new tasks by moving its
arms in the desired sequence
78. 78
History of Artificial Intelligence
Deep Learning
1998: Yann LeCun 's second
generation Convolutional Neural
Networks
2006: Geoffrey Hinton’s Geoffrey
Hinton's Deep Belief Networks
2007: Yeshua Bengio's Stacked
Auto-Encoders
Deep
Learning
79. 79
History of Artificial Intelligence
Deep Learning mimics the workings of the
brain: the audiovisual cortex works in
multiple hierarchical stages
80. 80
History of Artificial Intelligence
2014: Vladimir Veselov's and Eugene Demchenko's program
Eugene Goostman, which simulates a 13-year-old Ukrainian
boy, passes the Turing test at the Royal Society in London
2014: Li Fei-Fei's computer vision algorithm that can describe
photos ("Deep Visual-Semantic Alignments for Generating
Image Descriptions", 2014)
2014: Alex Graves, Greg Wayne and Ivo Danihelka publish a paper
on "Neural Turing Machines"
2014: Jason Weston, Sumit Chopra and Antoine Bordes publish a
paper on "Memory Networks"
2014: Microsoft demonstrates a real-time spoken language
translation system
2014: GoogLeNet, a 22 layers convolutional network, wins
Imagenet 2014, almost doubling the mean average precision of
the previous year’s winner
81. 81
Break
"There is more stupidity than hydrogen in the universe,
and it has a longer shelf life"
(Frank Zappa)
83. 83
Knowledge-based Systems
“General problem solver”:
the program capable of
solving all problems
Intelligence =
reasoning about
knowledge
Domain knowledge
and domain experts Knowledge
Representation
Knowledge-based
systems
(expert systems)
Knowledge
Engineering
84. 84
Knowledge-based Systems
• Knowledge representation
– Predicates
– Production rules
– Semantic networks
– Frames
• Inference engine: reasoning mechanisms
• Common sense & heuristics
• Uncertainty
• Learning
89. 89
Artificial Intelligence
A New Class of
Applications
Expert
Tasks
Heuristics Uncertainty
“Complex”
Problem Solving
The algorithm
does not exist
A medical
encyclopedia
is not
equivalent to a
physician
The algorithm is
too
complicated
Design a cruise
ship
There is an
algorithm
but it is
“useless”
Don’t touch
boiling water
The algorithm is
not possible
Italy will win the
next world
cup
90. 90
Artificial Intelligence
A New Class of
Applications
Expert
Tasks
Heuristics Uncertainty
“Complex”
Problem Solving
Expert
Systems
Vision
Speech
Natural
Language
91. 91
Artificial Intelligence
A New Class of
Technologies
Non-sequential
Programming
Symbolic
Processing
Knowledge
Engineering
Uncertain
Reasoning
94. 94
Expert System Manufacturing Cycle
Knowledge Acquisition
(Identify sources of expertise)
Knowledge Representation
(Define the structure of the
knowledge base)
Control Strategy
(Define the structure of the
inference engine)
Rapid Prototyping
(Generate and test)
Fine-tuning
(Evaluate feedback
from field)
95. 95
The Evolution of Computers in the
Information Age
Computers
Humans
Decision
Making
Data
Processing
1960 2014
Algorithmic Programming Artificial Intelligence
96. 96
2000s
• Computational power becomes a distraction
– Translation
– Search
– Voice recognition
based on statistical analysis, not “intelligence”
• Emphasis on guided machine learning, in most cases
probabilistic analysis of cases
• “Best Guess AI”
97. 97
Common Sense
Small minds are concerned with the extraordinary,
great minds with the ordinary"
(Blaise Pascal)
98. 98
Common Sense
• In which cases it is perfectly logical to find a
woman in the men’s restrooms?
100. 100
Common Sense
• How easy is it to build a robot that can make these
trivial inferences?
101. 101
Common Sense
• What is wrong about this picture?
Communication tower
on top of Mt Diablo (2015)
102. 102
Common Sense
• Deduction is a method of exact inference
(classical logic)
– All Greeks are humans and Socrates is a
Greek, therefore Socrates is a human
• Induction infers generalizations from a set of
events (science)
– Water boils at 100 degrees
• Abduction infers plausible causes of an effect
(medicine)
– You have the symptoms of a flue
103. 103
Common Sense
Types of inference
Induction
Concept formation
Probabilistic reasoning
Abduction
Diagnosis/troubleshooting
Scientific theories
Analogy
Transformation
Derivation
104. 104
Common Sense
• Classical Logic is inadequate for ordinary life
• Intuitionism (Luitzen Brouwer, 1925)
– “My name is Piero Scaruffi or 1=2”
– “Every unicorn is an eagle”
– Only “constructable” objects are legitimate
• Frederick and Barbara Hayes-Roth (1985):
opportunistic reasoning
– Reasoning is a cooperative process
carried out by a community of agents, each
specialized in processing a type of
knowledge
105. 105
Common Sense
• Multi-valued Logic (Jan Lukasiewicz, 1920)
– Trinary logic: adds “possible” to “true” and
“false”
– Or any number of truth values
– A logic with more than “true” and “false” is not
as “exact” as classical Logic, but it has a higher
expressive power
• Plausible reasoning
– Quick, efficient response to problems when an
exact solution is not necessary
• Non- monotonic Logic
– Second thoughts: inferences are made
provisionally and can be withdrawn at any time
106. 106
Common Sense
The Frame Problem
– Classical logic deducts all that is possible from all
that is available
– In the real world the amount of information that is
available is infinite
– It is not possible to represent what does “not”
change in the universe as a result of an action
("ramification problem“)
– Infinite things change, because one can go into
greater and greater detail of description
– The number of preconditions to the execution of any
action is also infinite, as the number of things that
can go wrong is infinite ("qualification problem“)
107. 107
Common Sense
Uncertainty
“Maybe i will go shopping”
“I almost won the game”
“This cherry is red”
“piero is an idiot”
Probability
Probability measures "how often" an event occurs
But we interpret probability as “belief”
Glenn Shafer’s and Stuart Dempster’s “Theory of
Evidence” (1968)
108. 108
Common sense
Principle of Incompatibility (Pierre Duhem)
The certainty that a proposition is true decreases
with any increase of its precision
The power of a vague assertion rests in its being
vague (“I am not tall”)
A very precise assertion is almost never certain (“I
am 1.71cm tall)
109. 109
Common Sense
Fuzzy Logic
Not just zero and one, true and false
Things can belong to more than one category, and
they can even belong to opposite categories, and
that they can belong to a category only partially
The degree of “membership” can assume any value
between zero and one
110. 110
Common Sense
The world of objects
Pat Hayes’ measure space (1978)
• measure space for people’s height: the set of natural
numbers from 100 (cms) to 200 (cms)
• measure space for driving speed: the set of numbers
from 0 to 160 (km/h)
• measure space for a shirt’s size: small, medium, large,
very large
John McCarthy's “Situation Calculus” (1963)
Qualitative reasoning (Benjamin Kuipers; Johan DeKleer;
Kenneth Forbus)
• Qualitative descriptions capture the essential aspects of
structure, function and behavior, e.g. “landmark” values
111. 111
Common Sense
Heuristics
• Knowledge that humans tend to share in a
natural way: rain is wet, lions are dangerous,
most politicians are crooks, carpets get
stained…
• Rules of thumbs
György Polya (1940s): “Heuretics“ - the nature,
power and behavior of heuristics: where it comes
from, how it becomes so convincing, how it
changes
112. 112
Common Sense
Douglas Lenat (1990):
• A global ontology of common knowledge and a set of
first principles (or reasoning methods) to work with it
• Units of knowledge for common sense are units of
"reality by consensus“
• Principle of economy of communications: minimize the
acts of communication and maximize the information
that is transmitted.
113. 113
Connectionism
"Politics is the art of looking for trouble, finding it
everywhere, diagnosing it incorrectly and applying the
wrong remedies"
(Groucho Marx)
114. 114
Connectionism
A neural network is a set of interconnected neurons
(simple processing units)
Each neuron receives signals from other neurons and
sends an output to other neurons
The signals are “amplified” by the “strength” of the
connection
115. 115
Connectionism
The strength of the connection changes over time
according to a feedback mechanism (output desired
minus actual output)
The net can be “trained”
Output
Correction
algorithm
117. 117
Connectionism
Where are we?
Largest neural network:
– 1,000,000,000,000 neurons
Worm’s brain:
– 1,000 neurons
But the worm’s brain outperforms neural networks on most
tasks
Human brain:
– 100,000,000,000 neurons
– 200,000,000,000,000 connections
118. 118
More Than Intelligence
• Summary
– Common Sense
– Deduction/ Induction/ Abduction
– Plausible Reasoning
– The Frame Problem
– Uncertainty, Probability, Fuzzy Logic
– Neural Networks
– Fault-tolerant
– Recognition tasks
– Learning
119. 119
Artificial Intelligence
• Notes…
– How many people can fly and land upside
down on a ceiling? Don’t underestimate the
brain of a fly.
– Computers don’t grow up. Humans do.
120. 120
Artificial Life
• 1947: John Von Neumann’s self-replicating and
evolving systems
• 1962: first computer viruses
• 1970: John Conway: The Game of Life
• 1975: John Holland’s genetic algorithms
• 1976: Richard Dawkins: the meme
• 1987: Chris Langton: "Artificial Life"
• 1988: Robert Morris: the first computer virus on
the Arpanet
• 1991: Thomas Ray: "Evolution and optimization
of digital organisms"
121. 121
Emergent Computation
• Alan Turing's reaction-diffusion theory of
pattern formation
• Von Neumann's cellular automata
– Self-reproducing patterns in a simplified
two-dimensional world
– A Turing-type machine that can reproduce
itself could be simulated by using a 29-state
cell component
1. Machine P (parent) reads the blueprint and makes a copy of itself, machine C (child).
2. Machine P now puts its blueprint in the photocopier, making a copy of the blueprint.
3. Machine P hands the copy of the blueprint to machine C.
The blueprint is both active instructions and passive data.
122. 122
Emergent Computation
• Turing proved that there exists a universal computing
machine
• Von Neumann proved that
– There exists a universal computing machine which,
given a description of an automaton, will construct
a copy of it
– There exists a universal computing machine which,
given a description of a universal computing
machine, will construct a copy of it
– There exists a universal computing machine which,
given a description of itself, will construct a copy of
itself
123. 123
Artificial Life
• John Holland's genetic algorithms (1975)
• Genetic algorithms as "search algorithms based
on the mechanics of natural selection and
natural genetics"
– “Reproduction“ (copies chromosomes
according to a fitness function)
– “Crossover“ (that switches segments of two
chromosomes)
– “Mutation"
– Etc
• Thomas Ray’s “Tierra” (1992)
124. 124
Artificial Life
• Possible solutions "evolve" in that domain until they
fit the problem.
• Solutions evolve in populations according to a set
of "genetic" algorithms that mimic biological
evolution.
• Each generation of solutions, as obtained by
applying those algorithms to the previous
generation, is better "adapted" to the problem at
hand.
125. 125
Artificial Life
• Frank Tipler: one has no way to tell a computer
simulation of the real world from the real world, as
long as one is inside the simulation
• The distinction between reality and simulation is
fictitious
• Artificial life replaces the "problem solver" of artificial
intelligence with an evolving population of problem
solvers
• The “intelligence” required to solve a problem is not
in an individual, it is in an entire population and its
successive generations
126. 126
Virtual Reality
• Morton Heilig: the Sensorama (1962)
• Rainer Werner Fassbinder: "Welt am Draht/ World on
a Wire" movie (1973)
• William Gibson: the cyberspace (1982)
• John Badham: "War Games" movie (1983)
• Bruce Bethke: "Cyberpunk" (1983)
• NASA Ames: Virtual Planetary Exploration
Workstation (1984)
• Lucasfilm: online game "Habitat" (1985) that
introduces "avatars"
• Neal Stephenson: the metaverse (1992)
• Philip Rosedale: Second Life (2003)
127. 127
Virtual Reality
• David Deutsch (1997)
– The technology of virtual reality (the ability of a
computer to simulate a world) is the very
technology of life
– Genes embody knowledge about their ecological
niche
– An organism is a virtual-reality rendering of the
genes
128. 128
Summary of the summaries
• David Hilbert (1928)
• The Turing Machine (1936)
• Cybernetics (1947)
• Information Theory (1949)
• Artificial Intelligence: Symbol Processing
• Artificial Intelligence: Connectionism
• Artificial Intelligence: Robots
• Common Sense
• Plausible Reasoning
• Recognition tasks
• Artificial Life